1
00:00:11,070 --> 00:00:16,530
OK, so in this next section of the course, we are going to look at what I think is an exciting opportunity

2
00:00:16,860 --> 00:00:20,030
to use industrial strength time series analysis.

3
00:00:20,580 --> 00:00:25,140
Now, usually when we do machine learning, we think about building the model ourselves.

4
00:00:25,320 --> 00:00:28,170
This is what most data scientists in the industry do.

5
00:00:28,830 --> 00:00:34,830
But what if you don't have data science or machine learning expertise on your team or what if you do?

6
00:00:34,830 --> 00:00:37,290
But they don't have experience in this domain?

7
00:00:37,920 --> 00:00:41,310
Well, luckily, there are services out there that you can use.

8
00:00:41,700 --> 00:00:46,260
This section is all about Amazon Web Services, also known as eight of us.

9
00:00:50,970 --> 00:00:56,610
So if you've taken my courses before, you already know that eight of us is a behemoth when it comes

10
00:00:56,610 --> 00:01:03,150
to cloud computing, we've used Adewusi to run large scale, mass produced jobs to create powerful GPU

11
00:01:03,150 --> 00:01:06,170
instances for training deep neural networks and more.

12
00:01:06,660 --> 00:01:11,820
Now, you might think of eight of us as just a service that provides you with virtual machines so that

13
00:01:11,820 --> 00:01:16,160
you can run websites or train machine learning models or do big data processing.

14
00:01:16,710 --> 00:01:19,340
But in fact, Adewusi does a lot more.

15
00:01:19,830 --> 00:01:23,520
In particular, it offers pre-built machine learning models.

16
00:01:23,940 --> 00:01:29,010
I'd encourage you to check out their services for yourself so you can see the wide range of applications

17
00:01:29,010 --> 00:01:36,090
they offer, from face detection to speech recognition to sentiment analysis to neural machine translation

18
00:01:36,330 --> 00:01:38,500
and yes, time series forecasting.

19
00:01:39,420 --> 00:01:44,550
So this section is all about how to do Time series forecasting with Adewusi.

20
00:01:49,130 --> 00:01:55,370
Now, some machine learning services on Adewusi offer you pre trained models, so all you have to do

21
00:01:55,370 --> 00:01:57,690
is call a function and you get an answer.

22
00:01:58,190 --> 00:02:03,530
For example, if you want to do sentiment analysis, you can just call a function in the Boto three

23
00:02:03,530 --> 00:02:09,350
library, passing your text and get a response, telling you what the sentiment of that text is.

24
00:02:09,950 --> 00:02:15,290
As a quick quiz question, let's think about why that would not be the case for a time series forecasting.

25
00:02:25,000 --> 00:02:30,280
OK, so hopefully you thought about why times for his forecasting isn't as simple as just calling a

26
00:02:30,280 --> 00:02:31,780
function in Boto three.

27
00:02:32,560 --> 00:02:37,550
The answer is, of course, that the forecast for your Time series depends on your Time series.

28
00:02:37,930 --> 00:02:41,790
Of course, Amazon hasn't necessarily seen your Time series before.

29
00:02:42,280 --> 00:02:48,160
This is unlike something like sentiment analysis where your text is probably very much like other text.

30
00:02:48,190 --> 00:02:49,240
Amazon has seen.

31
00:02:49,900 --> 00:02:54,700
So in order to make a forecast on your data, you have to actually train a new model.

32
00:02:55,300 --> 00:02:58,240
As you'll see, this is quite an involved process.

33
00:02:59,020 --> 00:03:02,710
That's why this is an entire section of a course and not just a lecture.

34
00:03:03,220 --> 00:03:07,080
In fact, I would say it's more difficult than what we did previously in the course.

35
00:03:07,570 --> 00:03:13,780
The upside is that you do get to work with a powerful industrial strength time series forecasting product.

36
00:03:18,470 --> 00:03:20,660
So who might be interested in this section?

37
00:03:21,290 --> 00:03:24,380
Well, let's say your boss asks you to use Amazon forecast.

38
00:03:24,860 --> 00:03:30,650
Suppose that the senior leadership at your company feels more comfortable using Amazon's solution instead

39
00:03:30,650 --> 00:03:32,090
of a custom made solution.

40
00:03:32,780 --> 00:03:38,120
Another reason you might want to use Amazon forecasts is that you've taken this course and you've realized

41
00:03:38,120 --> 00:03:41,440
that data science and machine learning is way over your head.

42
00:03:41,930 --> 00:03:47,090
You might be a software developer who's not used to writing algorithms from scratch, but rather using

43
00:03:47,090 --> 00:03:50,030
enterprise products purchased from other companies.

44
00:03:50,480 --> 00:03:53,690
In this case, you would feel right at home with Amazon forecast.

45
00:03:54,890 --> 00:03:59,990
Another reason you might want to use Amazon forecast is that you've tried everything, but your model

46
00:03:59,990 --> 00:04:01,700
isn't as accurate as you like.

47
00:04:02,150 --> 00:04:05,620
You might want to try Amazon forecast to see if it can do better.

48
00:04:07,120 --> 00:04:13,390
Another reason to use Amazon forecast is that it does automate and hyper parameter optimization for

49
00:04:13,390 --> 00:04:16,960
you, therefore, you don't have to think about those things yourself.

50
00:04:17,470 --> 00:04:22,090
So if you don't want to worry about choosing the right model or searching for the right hyper parameters,

51
00:04:22,240 --> 00:04:26,100
then Amazon forecast offers to take that work off your hands.

52
00:04:30,790 --> 00:04:36,790
Now, there are also some reasons why you might not want to use Amazon forecast for one, this is not

53
00:04:36,790 --> 00:04:37,350
free.

54
00:04:37,750 --> 00:04:42,010
You do have to pay for that auto Amelle and that type of parameter optimization.

55
00:04:42,460 --> 00:04:45,460
Every time you make a forecast, you have to pay for that.

56
00:04:45,790 --> 00:04:48,520
Every time you train a model, you have to pay for that.

57
00:04:48,850 --> 00:04:51,850
When you store data in three, you have to pay for that.

58
00:04:52,600 --> 00:04:55,170
So computation and data storage costs money.

59
00:04:56,020 --> 00:05:01,270
Another reason you might not want to use Amazon forecast is that it can be quite a bit slower than what

60
00:05:01,270 --> 00:05:02,050
you're used to.

61
00:05:02,560 --> 00:05:07,230
You've already seen how long it takes to train a CNN or an aunt in or in a remote model.

62
00:05:07,660 --> 00:05:13,000
What you're about to see is that the whole process for building and training a model with Amazon forecasts

63
00:05:13,090 --> 00:05:14,600
is surprisingly slow.

64
00:05:15,790 --> 00:05:21,130
Yet another reason you might not want to use Amazon forecast is you're about to see how complicated

65
00:05:21,130 --> 00:05:22,140
it really is.

66
00:05:22,630 --> 00:05:26,680
This is not because the modeling we're doing is particularly complicated.

67
00:05:26,800 --> 00:05:29,760
It's because how the API works is complicated.

68
00:05:30,160 --> 00:05:34,800
So after seeing how to use it in this section, you may decide that it's not for you.

69
00:05:35,350 --> 00:05:37,700
But again, that's the tradeoff you have to make.

70
00:05:38,020 --> 00:05:44,380
Yes, the API is a bit more complex, but you do get a really powerful model with lots of complex features.

71
00:05:49,050 --> 00:05:51,130
So here's a quick outline for this section.

72
00:05:51,780 --> 00:05:55,260
Now, this course is mostly focused on the data science side of things.

73
00:05:55,740 --> 00:06:01,740
I'm going to assume a certain level of competence with programming, reading documentation, using Web

74
00:06:01,740 --> 00:06:03,150
services and so forth.

75
00:06:03,600 --> 00:06:09,510
So this will not be a total beginners introduction for how to sign up for an account with us, how to

76
00:06:09,510 --> 00:06:11,320
get your credentials and so forth.

77
00:06:11,700 --> 00:06:17,010
I'm going to assume that if you are interested in working with Amazon Web Services, that you have the

78
00:06:17,010 --> 00:06:21,420
capability of an average person who actually uses Amazon Web services.

79
00:06:22,870 --> 00:06:28,780
Also note that if you work at a company, the sign up process and your credentials will probably be

80
00:06:28,780 --> 00:06:29,680
handled for you.

81
00:06:30,250 --> 00:06:36,070
So once you actually have an account and you know how to use it, we'll start by creating an email specifically

82
00:06:36,070 --> 00:06:38,010
for use with Amazon forecast.

83
00:06:38,740 --> 00:06:39,190
I am.

84
00:06:39,190 --> 00:06:43,330
Rules are basically Amazon's way of handling security and authentication.

85
00:06:44,440 --> 00:06:46,690
Next, we'll look at how to upload our data to us.

86
00:06:46,690 --> 00:06:51,430
Three as three is another Amazon Web service use to store data.

87
00:06:52,300 --> 00:06:55,000
This is where a model will look for its training data.

88
00:06:55,780 --> 00:06:58,200
The next step will be to build and train our model.

89
00:06:58,810 --> 00:07:03,010
Once we've done that, we'll see how we can use our trained model to make forecasts.

90
00:07:04,370 --> 00:07:10,220
Finally, we'll look at how to evaluate how good our model is, one interesting thing about this process

91
00:07:10,220 --> 00:07:15,560
is that at a high level, you can see that it's pretty much exactly the same as the first section of

92
00:07:15,560 --> 00:07:16,310
this course.

93
00:07:16,730 --> 00:07:20,620
Get the data, build the model, train the model and evaluate the model.

94
00:07:21,200 --> 00:07:24,740
As you recall, that ended up being only a few lines of code.

95
00:07:25,220 --> 00:07:26,660
So what's different now?

96
00:07:27,170 --> 00:07:32,730
So although the high level steps are the same, the actual syntax in API is very different.

97
00:07:33,140 --> 00:07:34,350
So that's the tradeoff.

98
00:07:34,610 --> 00:07:40,340
Yes, you do get to use an industrial strength time series forecasting solution, which may or may not

99
00:07:40,340 --> 00:07:45,580
be better than just a basic Python library, but each of these steps take significantly more work.

100
00:07:50,320 --> 00:07:55,990
To prepare for this section, you'll need to install the Boto three library as usual, you can use the

101
00:07:55,990 --> 00:07:57,490
PIP install command.